The stuff about how 4.8 should be using same amount of tokens as 4.7 hints that Anthropic might be acknowledging how token-inefficient they are compared with GPT-5.5. I wonder how long it will take before the corporate customers openly complain about the inference costs of Opus as opposed to GPT.
As someone who uses Claude chat for coding, 4.7 was awful and I went back to 4.6 and will stay until they remove it.
4.6 itself has added more "reasoning" which is where it debates with itself about whether it should follow my instructions or make up total nonsense I said i dont want.
Reasoning may work for some cases, but its awful for keeping constraints because its all about being less confident and second guessing, making all results worse. More LLM is worse LLM. I hope I can finish my current big project before they make it unusable for my purposes.
Whew. I barely followed any of that... but grasped some sense of the flow (expansion and honing?) of the developing and refining of the work. And my angry cynical mind threw a rotten tomato at Trump saying "we don't HAVE enough Americans capable of that level of work."
Wanna BET the majority of the programmers, developers, and planners are WHITE MEN!?
The stuff about how 4.8 should be using same amount of tokens as 4.7 hints that Anthropic might be acknowledging how token-inefficient they are compared with GPT-5.5. I wonder how long it will take before the corporate customers openly complain about the inference costs of Opus as opposed to GPT.
Unfortunately, 4.8 can't write fiction either. It's terrible in a different way than 4.7 was.
What do you use for translating fiction? That seems a more constrained task. Does the problem apply in this case too?
4.6 and 4.8 both work fine. The problem doesn't apply because it has a strong anchor for every word. So it doesn't have to invent anything.
As someone who uses Claude chat for coding, 4.7 was awful and I went back to 4.6 and will stay until they remove it.
4.6 itself has added more "reasoning" which is where it debates with itself about whether it should follow my instructions or make up total nonsense I said i dont want.
Reasoning may work for some cases, but its awful for keeping constraints because its all about being less confident and second guessing, making all results worse. More LLM is worse LLM. I hope I can finish my current big project before they make it unusable for my purposes.
That's a great summary: debates with itself about whether it should follow my instructions or make up total nonsense I said i dont want.
Whew. I barely followed any of that... but grasped some sense of the flow (expansion and honing?) of the developing and refining of the work. And my angry cynical mind threw a rotten tomato at Trump saying "we don't HAVE enough Americans capable of that level of work."
Wanna BET the majority of the programmers, developers, and planners are WHITE MEN!?